We propose physics-based methods that will assist in comparative protein structure modeling. Although ab initio methods won't soon replace comparative modeling, particularly for large proteins, multiple domains, or genome-wide studies, we believe that physics-based methods are now poised to play a big role in improving comparative models. By improving the efficiency of conformational sampling and the accuracy of atomistic energy functions, we propose ways to circumvent alignment problems where sequence identity is low, and to give atom-level refinements where sequence identity is high. Our aims are aligned with both comparative modeling goals identified in RFA-GM-07-003, "High Accuracy Protein Structure Modeling." [unreadable] [unreadable] (1) Better sampling using kinematics. To improve the modeling of concerted motions in constrained structures, like loops and helices, we will use analytical kinematics methods that we have recently developed, resembling those used in robotics of systems of linked rods. We will combine the kinematics with a powerful multiple temperature scheme (replica exchange) to further increase sampling efficiency. We will apply these methods primarily to homology models having good sequence alignments (typically >30% sequence identity). [unreadable] [unreadable] (2) Better sampling using "Zipping an Assembly" with bioinformatics restraints. The advance here is a new highly efficient protein-folding mechanism-based search method called zipping and assembly, which is recently CASP-tested. Two key features of this approach are that it should: (a) tolerate sequence alignment errors, and (b) provide models for large insertions, which are not aligned to template residues. Our goal here is to help remote comparative modeling. [unreadable] [unreadable] (3) Better atomistic energy-based scoring functions. The end game in protein structure prediction requires scoring functions that are correct in detail, hence correct in the physics. We propose two ways to improve them. First, we will improve a key defect in implicit solvent models by including a better treatment of the first shell of water around the molecule. Second, we have previously developed a general approach to parameter optimization, called MOPED, which is applicable to complex nonlinear multi-parameter models. We will apply it to improving parameters, using the large datasets generated in aims 1 and 2. [unreadable] [unreadable] We will test our approaches in blind predictions including CASP. We will make our results freely available through modular source codes and executable programs. [unreadable] [unreadable] [unreadable]